AddressHate Annotator

Annotate and flag hate speech where you find it, on the platforms AddressHate studies. Sign in to work with your research team.

AddressHate Annotator is the labelling tool used by the AddressHate research team to document online hate speech. It lets an authorised researcher record a structured judgement about a post without leaving the page it appears on, and files that judgement in the team's own annotation service. It is a research instrument rather than a consumer product, and what an account may do depends on who granted it. Anyone may create one and confirm their address; that account can report material it comes across, which joins a queue for the teams who study it. Labelling to a study's own codebook is a further permission, given by the team running that study — by inviting the account, or by approving a request from a recognised academic address. Signing in works with a work or school account, or with a password. WHERE IT RUNS The extension activates only on the specific social platforms and AI assistant sites this project studies. The full list is shown in this item's site-access details. Everywhere else it is inert: it does not read, alter or transmit anything from any other page. LABELLING IN PLACE On a supported site the extension adds an Annotate control to individual posts, comments, videos and conversation turns, and marks the ones already covered so a team never does the same work twice. Opening the drawer, a researcher answers the questions their study asks — a severity judgement, whichever conceptual categories apply, written notes, ratings, or a passage highlighted in the text itself. Nearby replies can be pulled in as context when something only makes sense beside what it responds to. The same drawer can also be opened from the extension's own workspace, over an item a lead has queued, so a list of collected material can be worked straight through with a link out to the original wherever the pictures and the surrounding conversation matter. QUESTION SETS THAT MATCH THE STUDY Different studies ask different things. An administrator can build a question set in the browser — field types, option lists, follow-up questions that appear only when an earlier answer warrants them, and an optional custom layout — then point a team or a single project at it. The drawer renders exactly that instrument. Every answer is stored against the version of the question set it was given under, so re-wording a question later never quietly changes what someone meant months ago. ORGANISING THE WORK Work is grouped into projects. A lead can queue a list of links as tasks, hand them to particular people, and watch them tick off as they are finished. A researcher chooses the project they are working through, and everything they label is filed under it — whether it came from the queue or was found while reading. Projects nest, so a large study divides into batches without losing the thread. REVIEW BEFORE ANYTHING COUNTS Nothing is treated as finished on its author's say-so. Each submission waits for a lead, who either signs it off or asks for changes with a note saying what to revisit. Where two people have judged the same item differently, a dedicated view sets their answers side by side so the disagreement can be resolved and the codebook sharpened. Everyone can see the state of what they have submitted, along with any feedback left for them. GETTING THE DATA OUT A finished body of work exports to a spreadsheet whose columns follow the question set it was collected under: one sheet per instrument, one column per question, and answers written out as the wording that was chosen rather than as internal codes. The material still waiting to be worked on exports the same way. The same download can be taken as structured data instead, put on a repeating delivery to a mailbox or a storage bucket, or fetched by a script through a read-only key that a lead issues and can withdraw at any time. USABLE BY EVERYONE The interface targets WCAG 2.1 AA. Controls are reachable and operable from the keyboard, focus is always visible, the layout reflows without sideways scrolling, and colour is never the only carrier of meaning. An automated audit runs against the real interface on every build. WHAT IS COLLECTED There is one destination for research material: the project's own service. Nothing is sent at all until a signed-in researcher deliberately labels or reports something. Pages that are merely read are not gathered, not logged and not transmitted. Two things involve anyone else, and both are named here rather than left to be found. When a researcher asks for an automated second opinion on a post, that post's text is sent to the AI provider the project uses, and no account details travel with it. And when an account is set up, an unticked box — never required for any part of the tool — offers to add that address to the project's mailing list, which runs on a newsletter service: ticking it sends the address, the name on the account and the organization it belongs to, and leaving it alone sends nothing anywhere. Every message carries an unsubscribe link. What is kept, why, and how to have it deleted is set out in full on the Privacy Policy and Data Deletion pages linked from this listing. REPORTING A PRIVATE CONVERSATION One case works differently and deserves stating plainly. Something posted publicly can be reported by link alone, because a reviewer can open that link and see it. An exchange with an AI assistant is private to the account that held it, so a link would show a reviewer nothing at all. Reporting one therefore captures the reply being raised, the prompt that produced it, and a short run of the turns before it. Beforehand the extension displays exactly what it has gathered and lets the researcher strike out the surrounding turns; cancel, and nothing whatsoever leaves the machine. WHY IT EXISTS AddressHate is a non-profit research effort. This tool exists to measure and document hate speech so that it can be understood and countered — never to create, promote or amplify it.